Is Claude Opus 5.5 nerfed? What the trackers show, and what to check

What people are saying

At 4:15 p.m. Central on November 7, 2026, Shane Melaugh posted on X: “is Opus 5.5 nerfed? I just had Gemini 4 Argon analyze 1,400 Claude Code sessions and it found that on November 3, something definitely happened. the model name didn’t change, but the way it does reasoning took a sharp turn.” In replies he wrote that “how much reasoning Opus 5.5 does in ‘high’ took a nosedive,” that on a real bug “Opus 5.5 xhigh takes a moment to think, finds a fix after a few minutes” where high didn’t, and: “so yeah, you’re not imagining things. the nerf is real.”

AICoder counted about 113,000 impressions of it. It also reports that the developer Theo answered that Anthropic appeared to be rolling out a change to its reasoning summarizer, account by account, and that he saw no sign of weaker capability. AICoder paraphrases him and doesn’t link his post. Melaugh’s analysis is one person’s sessions, judged by another company’s model, and he hasn’t published the data.

It’s the latest round of a complaint as old as the model. Opus 5.5 came out on September 22, and on October 5 a GitHub issue said its quality had “degraded back to level of opus 5” (#99572). The most detailed report, #105112, opened October 26, comes from a user who scored their own transcripts: since October 19, “Opus 5.5’s answers in my Claude Code sessions have been corrected about five times as often as in its first four weeks: 25.4% of 900 answers, against 5.2% of 5,738.” Their update for November 4 reads “still about four times the pre-Oct 19 error rate, more than two weeks on.” A Max 20x subscriber working at high effort filed #105477 the next day. These are users’ own measurements, and neither issue has a reply from Anthropic.

What the trackers measure

Three projects rerun the same tests on Opus 5.5 to catch exactly this. None has found a decline:

Tracker	What it runs	Latest reading	Its verdict
LiveNerf	78 fixed questions from GPQA Diamond, MMLU-Pro, competition math and AIME, daily, through Claude Code 2.1.280 on a Max plan	0.614 for October 24 to November 2, against 0.584 at launch	“no change” in all three windows
AI Stupid Level	Coding, reasoning and tool-use tests	95 ± 2.3 out of 100, retested November 9	“no sustained decline detected”
BridgeBench NerfBench	Coding and reasoning tasks, scored against its launch run	105.8% of launch on October 16	“No nerfs detected”

Two limits matter. First, they test fixed, mostly short tasks, while the loudest complaints come from long agent sessions, a gap MODEL points out: both can be true. Second, timing. LiveNerf pins an older Claude Code, so it measures the model rather than the app, and its window that began on November 3 has no result until it closes on November 12. Of the three, only AI Stupid Level has a published reading we could find from after Melaugh’s date.

What Anthropic has said

Nothing about these complaints, as of November 9. Its status page lists no Opus 5.5 incident since October 24, and none on November 3, though it tracks errors rather than quality. Its models overview says “Every Claude model ID is a pinned snapshot,” so the model behind claude-opus-5-5 isn’t swapped under the same name.

Anthropic has answered the question before. In September 2025 it wrote: “To state it plainly: We never reduce model quality due to demand, time of day, or server load.” In its April 23, 2026 postmortem it wrote that “We never intentionally degrade our models,” and then traced weeks of complaints about Claude Code to three changes around the model: a lower default effort, a caching bug that cleared Claude’s earlier thinking, and a prompt line that limited verbosity. All three were fixed by version 2.1.116. A model can stay the same while the app around it changes.

One documented case does put another model in your conversation. Opus 5.5’s safeguards “fall back to another model transparently,” its launch post says, and “most cybersecurity tasks will be re-routed to Opus 4.8.” Anthropic says Claude shows a notice when it happens; our memo on the fallback covers it.

What changed on November 3

One thing is on the record. Claude Code 2.1.317, released November 3 according to its changelog, “Changed subagents and teammates on a different model from the main conversation to use that model’s saved effort level or its default, instead of the session’s effort level.” If your session runs at high and a subagent runs on Sonnet 5.5 or Haiku 5.5, that subagent now starts at medium, their default in Claude Code, unless you’ve chosen a level for that model or its definition sets one. The main conversation isn’t affected.

The reasoning people read isn’t the reasoning itself. Anthropic’s thinking docs say “the thinking text you receive is a summary of Claude’s full thinking process rather than the raw chain of thought,” and “Summarization is processed by a different model from the one you target in your requests.” A change to that summarizer, as Theo suggests, could change how 1,400 sessions read to another model without any change to Opus 5.5. Anthropic hasn’t confirmed that either.

The setting behind a lot of it: effort

Opus 5.5 thinks less by default than the Opus before it. From Anthropic’s effort docs: “Claude Opus 5.5 supports all five effort levels, and medium is the default (Claude Opus 5 and earlier Opus models default to high, so a request that omits effort runs one level lower than it did on Claude Opus 5).”

Model	Default effort on the Claude API	Default effort in Claude Code
Opus 5.5	Medium	Medium
Sonnet 5.5	High	Medium
Haiku 5.5	Medium	Medium
Opus 5	High	High
Fable 5.1	High	High

Claude Code adds a catch. Its model configuration docs say “a top-level effortLevel in your user settings file doesn’t count for Opus 5.5”: if you set high with /effort for an earlier model, Opus 5.5 still starts at medium “until you choose a level for them with /effort or the /model picker.”

Anthropic’s case for medium is in its launch post: “Opus 5.5 at default effort beats Opus 5 at max effort for about a fifth of the cost.” But the post’s headline scores weren’t at the default: “Unless otherwise noted, all Claude Opus 5.5 results use adaptive thinking at max effort.” And in April, Anthropic called its earlier move of Claude Code’s default from high to medium “the wrong tradeoff.” On AI Coding Daily’s coding test through Claude Code in October, Opus 5.5 scored 67.26 of 70 at high, 65.23 at medium and 60.10 at low.

What to do if Opus 5.5 feels worse

1. Check your effort. In Claude Code, /effort opens a slider and /effort high or /effort xhigh sets the level. Enter saves it as Opus 5.5’s default; s in the slider keeps it to this session. The effort docs suggest high for “Complex reasoning, difficult coding problems, agentic tasks” and xhigh for “Long-running agentic and coding tasks (over 30 minutes) with token budgets in the millions.”
2. Raise it for one hard turn. Put ultrathink in a prompt to get “deeper reasoning on that turn without changing your session effort setting.” Claude Code’s docs warn that max “may show diminishing returns and is prone to overthinking.”
3. Set your subagents’ effort. Since 2.1.317, add effort to a subagent’s definition, or set CLAUDE_CODE_SUBAGENT_EFFORT, added in 2.1.316.
4. Check for a switch. A notice that Claude changed models means another model wrote that answer.
5. Compare like with like. Rerun a task you saved from launch week at the same effort and Claude Code version, rather than against memory. The author of #105112 found fresh sessions slip too, so a new session alone may not fix it.
6. Report it with numbers. Use /feedback in Claude Code, or add your version, effort level and dates to the GitHub issues above.
7. Get a second opinion. Give the same task to another lab’s agent. In Enjoy, Claude Code also starts at Medium; you can choose High and make it the default for a project, or switch the conversation to Codex, which catches up on what happened so far. Each agent runs on your own plan, and Enjoy doesn’t change how Anthropic serves the model.

Higher effort costs more: more tokens on the API, more of your plan on a subscription. If a higher level drains your limits, our guide to Claude Code’s usage limit has the options.

What Anthropic hasn’t said

• Whether anything changed in how claude-opus-5-5 was served on October 19 or November 3, the dates #105112 and Melaugh point to.
• Whether it changed how thinking summaries are written, as Theo suggests.
• Anything in reply to #105112 or #105477, as of November 9.

We’ll update this memo if Anthropic responds, or when LiveNerf’s window that began on November 3 closes on November 12.